Skip to content

WeeklyTelcon_20180717

Jump to bottom
Geoffrey Paulsen edited this page Jan 15, 2019 · 1 revision

Open MPI Weekly Telecon

  • Dialup Info: (Do not post to public mailing list or public wiki)

Attendees

  • Jeff Squyres
  • Geoff Paulsen
  • Josh Hursey
  • Brian
  • akshay
  • Edgar Gabriel
  • Geoffroy Vallee
  • Matias Cabral
  • Nathan Hjelm
  • Peter Gottesman (Cisco)
  • Todd Kordenbrock
  • Xin Zhao
  • Howard.
  • Joshua Ladd

not there today (I keep this for easy cut-n-paste for future notes)

  • Ralph Castain
  • Thomas Naughton
  • Akvenkatesh (nVidia)
  • Matthew Dosanjh
  • Howard Pritchard
  • Dan Topa (LANL)
  • David Bernholdt
  • Dan Topa (LANL)

Agenda/New Business

  • default to external for v4.0

    • should this be a blanket statement, or is there a version limit? For example, if someone has a v1.1 version of PMIx installed in a default location, do we really want to use that versus the internal v3.0?
      • if external support is found and is "compatible"... keep it fuzy.
        • If we know it won't work, just hardcode versioning.
        • If it's compatible, but lower than internal, we'd emit a warning, but still use external one.
      • no issue with PMIx v3.0 vs 2.x, since we haven't implemented new features anyway.
      • A little worried about PMIx, since older slrums are still pmix v1.x based.
      • if they're using pmix v1.2.5, ompi needs to use v1.2.5, since older pmix won't use newer pmix.
      • Does PMIx website have this compatibility chart somewhere? We should point to it.
      • We like checking at configure time, since it's a nice early failure, but this will have to be runtime failures.
      • configure summary at end. opal_summary_add m4.

  • Comm_Spawn issue:

    • George has raised a comm_spawn issue about inheriting MCA params https://github.com/open-mpi/ompi/issues/5376
      • Child job is launched using the same MCA params as first node. Isn't an easy way to overwrite those, and some can't be at all (other than don't use that).
      • He proposed mca parameters should not be inherited by child jobs. If it's set on command line, as opposed to mca param file, then it shouldn't be passed to child jobs.
    • for Open MPI v4.0
      • Just a couple of lines of code to NOT propagate the MCA params related to launch.
      • map-by option - can't turn off. (perhaps we need a sentinal meaning "default")

  • nVidia - update for UCX CUDA support

    • ompi_opal converter does not need any changes (verified by Akshay), changes thought needed, not needed.
    • See more on UCX and --with-cuda below.

  • github suggestion on email filtering

Minutes

Review All Open Blockers

Review v2.x Milestones v2.1.4

  • v2.1.4 - Final release on v2.1.x
  • Aug 10th is release date.
  • Targeting RC1 for v2.1.4 next week.
  • Will need PMIx fix from Issue 5336

Review v3.0.x Milestones v3.0.3

  • Schedule:
  • v3.0.3 - targeting Sept 1st
    • Anticipate RC1 after Aug 10th release of v2.1.4 releases.
    • Cisco is seeing some weirdness in v3.0 and v3.1
      • Haven't nailed down, but haven't reported yet. PMIx / runtime.
      • Ralph wants to see what happens when upgrade to PMIx v2.2, but probably a problem in ORTE.
  • Growing Pile of PRs that don't have reviews.

Review v3.1.x Milestones v3.1.0

  • v3.1.2 release process, starts after Sept 1st release of v3.0.3
  • Lots of PRs
  • Power9 hang in make check
  • Issue 5363 - teardown in shmem that trying to ulink a file already ulinked.
    • Fixed in master Cherry picked to other branches.
  • Issue 5336 - Brian merged PR that will print a bit more info. PMIx + libevent issue hit in cisco.
    • Giles tracked it down last night, and Ralph will PR it. Will need to go into v2.1.x also
  • PMIx v2.1.2 in Open MPI v3.1 and v3.0 - Ralph about to release PMIx v2.1.2 does Open MPI want to embed it?
    • All just bug fixes, not much work for Ralph to update.
    • Done.

v4.0.0

  • Schedule: branch: July 18. release: Sept 17

    • Date for all MTT testing online - July 22? -
    • Date for first RC - Aug 13 (after sunset of 2.1.4)

  • Targeting UCX v1.4 to support CUDA buffers.

    • May be changes in UCX PML and/or datatype converter.
    • Will have more info by next week.

  • Create the branch, then after that's created, then create nightly tarballs, and only after that can MTT be setup.

    • ACTION - send notes to brian to pre-emptively tell us if there will be issues.

  • Cuda support - cudasm, and openib

    • Still a couple of steps away of being on par in UCX regarding CUDA support.
    • Does nVidia want if --with-cuda, then openib included by default?
      • Yes, because at this moment UCX is not on par, but still want to migrate to ucx cuda.
      • Warning message will mention deficate openib vs ucx

  • NEWS - Depricate MPIR message for NEWs - Ralph can help with this.

  • Sent email to ompi-packagers list with schedule and info on

    • Debian (Allister)
    • Jeff has a half-typed out reply. Allister asked Do we really need to change the major .so version?
      • His point is that they have more and more packages compiled against Open MPI.
      • The real question is Do we mark the version as backwards incompatible?
      • The idea that you have to have the same version of Open MPI everywhere.
      • Nathan would say (watching what's gone in), that we should be okay, but that we should look.
    • In v3.1 --mpi-compat was on by default. In v4.0 there's a flag where the ABI didn't change.
      • MPI_UB was #define to &ompi_datatypeT
      • Could put the symbols back so they're there.
    • Fork in the road here... A couple of options:
      1. Set the CRA (.so versions) (Assuming --mpi-compat is NOT enabled). An app compiled against v3.1 won't run with v4.0
      2. Dynamicly set the CRA values (.so versions) based on the (--mpi-compat flag) bunch of maintence work.
        • Really HAIRY!
      4. NICE: Could Make --mpi-compat only affects the mpi.h, but don't affect the symbols in back end.
        • As long as they go away eventually - no point in removing from standard if we don't eventually remove them.
        • Raise C by 10, AND raise A by 10 - Make sure we get it right (do it like a minor release bump)
          • use rules from v3.0.x to v3.0.x+1 -
      • Don't know how this affects Fortran. (seperate MPI_UB and MPI_LB). It's the same as in C. Fail to compile without --mpi-compat
    • Symbols WILL go away in v5.0
    • Geoff and Howard will build test suites with v3.1.x and run with master/v4.0 to see if anything breaks.

  • Still at risk features for branch of v4.0 on July 15.

    • UTK ULFM - Fault Tollerant - Geoff JUST emailed George after meeting.
    • external preferences configury - Giles did libevent.
      • Ralph will update PMIx.
      • leaves hwloc - Jeff will review Giles code, and see if it can be easily translated to hwloc
        • potentially miss v4.0

PMIx

  • Ralph merged in some PMIx v3.0
  • Overall Runtime Discussion (talking v5.0 timeframe, 2019)
    • TODAY - Geoff Paulsen will send out doodle for next week to devel-core.
    • 10am Central July 18

New topics

  • From last week:
    • MTT License discussion - MTT needs to be de-GPL-ified.
    • Last week Brian had an interesting proposal to remove all of the perl out, or the python out?
    • Next week we'll Have Brian on the call.
    • Schedule - Like resolution by end of july.

Overall Runtime Discussion (talking v5.0 timeframe, 2019)

  • Will discuss this in a sperate call 2nd week in July.
  • Two Options:
    1. Keep going on our current path, and taking updates to ORTE, etc.
    2. Shuffle our code a bit (new ompi_rte framework merged with orte_pmix frame work moved down and renamed)
      • Opal used to be single process abstraction, but not as true anymore.
      • API of foo, looks pretty much like PMIx API.
        • Still have PMIx v2.0, PMI2 or other components (all retooled for new framework to use PMIx)
      • to call just call opal_foo.spawn(), etc then you get whatever component is underneath.
      • what about mpirun? Well, PRTE comes in, it's the server side of the PMIx stuff.
      • Could use their prun and wrap in a new mpirun wrapper
      • PRTE doesn't just replace ORTE. PRTE and OMPI layer don't really interact with each other, they both call the same OPAL layer (which contains PMIx, and other OPAL stuff).
        • prun has a lam-boot looking approach.
      • Build system about opal, etc. Code Shufflling, retooling of components.
      • We want to leverage the work the PMIx community is doing correctly.
  • If we do this, we still need people to do runtime work over in PRTE.
    • In some ways it might be harder to get resources from management for yet another project.
    • Nice to have a componentized interface, without moving runtime to a 3rd party project.
    • Need to think about it.
  • Concerns with working adding ORTE PMIx integration.
  • Want to know the state of SLURM PMIx Plugin with PMIx v3.x
    • It should build, and work with v3. They only implemented about 5 interfaces, and they haven't changed.
  • A few related to OMPIx project, talking about how much to contribute to this effort.
    • How to factor in requirements of OSHMEM (who use our runtimes), and already doing things to adapt.
    • Would be nice to support both groups with a straight forward component to handle both of these.
  • Thinking about how much effort this will be. and manage these tasks in a timely manor.
  • Testing, will need to discuss how to best test all of this.
  • ACTION: Lets go off and reflect and discuss at next week's Web-Ex.
    • We aren't going to do this before v4.0 branches in mid-July.
    • Need to be thinking about the Schedule, action items, and owners.

Review Master Master Pull Requests

  • Decided to file PR5200 to begin the long process of deleting osc/pt2pt (by enabling all relevant RDMA BTLs so that every transport will use osc/rdma).
  • Anything Jeff can help with Absoft and NAG licenses?
    • waiting.

Review Master MTT testing

  • Annual committer by July 18

    • AMD, ARM, Ructers, Mellanox, nVidia - Please go do Annual Commiter Reviews.

  • Hope to have better Cisco MTT in a week or two

    • Peter is going through, and he found a few failures, which some have been posted.
      • one-sided - nathan's looking at.
      • some more coming.
    • OSC_pt2pt will exclude yourself in a MT run.
      • One of Cisco MTTs runs with env to turn all MPI_Init to MPI_Thread_init (even though single threaded run).
        • Now that osc_pt2pt is ineligible, many tests fail.
        • on Master, this will fix itself 'soon'
        • BLOCKER for v4.0 for this work so we'll have vader and something for osc_pt2pt.
        • Probably an issue on v3.x also.

  • OSHMEM v1.4 - cleanup work

    • and refactoring.

  • Edgar has some issues running on omnipath - Not able to open HFI correctly.

    • Not sure if it's OFI components.
    • Mathias just updated his PR5004 and asked Jeff to review.
      • libfabric related, but probably not Edgar's issue.
    • Will retest current status.

  • Next Face to Face?

    • When? Late summer, early fall?
    • Where? San Jose - Cisco, Albuquerque - Sandia
    • Super computing is in Dallas this year in Nov.

Oldest PR

Oldest Issue

Status Updates:

Status Update Rotation

  1. Mellanox, Sandia, Intel
  2. LANL, Houston, IBM, Fujitsu
  3. Amazon,
  4. Cisco, ORNL, UTK, NVIDIA

Back to 2018 WeeklyTelcon-2018

Clone this wiki locally